Nondestructive perch target detection and size measurement from RGB-D images in recirculating aquaculture system

Abstract

In recirculating aquaculture system, precise estimation of perch size from images is essential for developing intelligent management system. However, variations in fish postures and visual field lead to different fish sizes in RGB images, posing challenges for accurate fish detection and localization. To address the above issues, this paper proposes a nondestructive target detection and size measurement method for perch, based on depth information and RGB-D images. The details are as follows: firstly, the capture ability of perch key point features is augmented by a bi-level routing attention (BRA) mechanism. Secondly, the enhanced CSPDarknet53 to 2-Stage FPN(C2f) module and new detection layer are introduced into the model’s backbone and neck, further improving the learning ability of perch features and the recognition accuracy of small-size targets. Finally, based on the detected key point coordinates, the perch size is calculated by combining the three-dimensional transformation from depth camera and measurement model. The experimental results show that the mAP@.5:.95 for key point detection reaches 86.4%, which is 3.6% higher than the baseline model, and the average relative error of perch size measurement is ± 5%. The proposed model provides an important basis for developing scientific feeding strategies and harvest plans.